cs.AIAug 25, 2025
SaveLLM-based Agentic Reasoning Frameworks: A Survey from Methods to Scenarios
Abstract
Recent advances in LLM-based agents highlight the importance of their reasoning frameworks, which guide the problem-solving process in diverse ways. This survey introduces a unified formal language to systematically categorize these frameworks at three compositional levels: single-agent, tool-based, and multi-agent methods. Following our taxonomy, we review key application scenarios across scientific discovery, healthcare, software engineering, society, economics, and general-purpose tasks. It also compares the distinct features and evaluation strategies of each category. Through our taxonomy and comparisons, our survey explores the designs and strengths of LLM-based agentic frameworks in different scenarios, reviewing the fast-paced development of complex agentic systems in the real world.
Figures & tables
Figure 1. Growth of indexed LLM-agent literature from Semantic Scholar Graph API 1 1 footnotemark: 1 keyword searches. Left: three framework queries. Right: four domain queries and a general-purpose agent query. Solid lines count matching journal and conference papers by publication year (2020–2025). Dashed red lines sum citations accrued by papers published in each year in the corresponding citation search, measured at retrieval. All queries were retrieved on 29 September 2026; recent counts may reflect indexing lag. Two line charts span 2020 to 2025. The left chart shows indexed publication counts for broad LLM-agent, multi-agent and tool-use framework keyword searches, with a separate citation-cohort curve on the right axis. The right chart shows indexed publication counts for scientific discovery, social and economics, software engineering, healthcare, and general-purpose agent keyword searches, with a separate citation-cohort curve on the right axis. In 2025 the five right-panel series respectively reach 209, 558, 473, 442 and 324 records.
Table 1. Coverage of related surveys published through 2025, ordered from older to newer, across methods, scenarios and evaluation. : focused coverage; : not a focus. The comparison criteria are defined in the text.
Figure 2. Taxonomy of agentic reasoning frameworks. We organize mechanisms at three compositional levels: a) single-agent methods, b) tool-based methods, and c) multi-agent methods. A framework can combine mechanisms across these levels. This image presents a detailed, hierarchical breakdown of agentic reasoning frameworks. The taxonomy is organized into three main categories: Single-agent Methods, which focus on prompt engineering and self-improvement techniques; Tool-based Methods, which cover how agents integrate, select, and utilize external tools; and Multi-agent Methods, which describe how multiple agents are organized and how they interact with each other. The diagram effectively visualizes the various sub-categories within each method, such as reflection, iterative optimization, and different forms of tool integration and agent interaction. This structure provides a clear and comprehensive overview of the different approaches used in designing agentic reasoning systems.
Figure 3. Prompt engineering for agentic reasoning frameworks. We summarize four types of methods: a) Role-playing : an agent will be distributed with a specific role, to stimulate its specific performance; b) Environmental-simulation : an agent will be told in a carefully designed environment, where it can reason with multi-modalities or external abilities; c) Task-description : a task will be clearly reconstructed and expressed to an agent; d) In-context Learning : several examples will be provided to an agent before or during its multi-step reasoning. For each type of prompting method, we provide a short example prompt, with the theme of conducting agentic research. This image is a detailed diagram illustrating four types of Prompt Engineering methods for agentic reasoning frameworks. The diagram breaks down each method with a short description and a corresponding example.
Figure 4. Paradigms of self-improvement for an LLM-based agent. We introduce three core mechanisms. a) Reflection : The agent analyzes a completed trajectory to generate a textual summary, storing in its context. This summary will helps for the next reasoning step. b) Iterative Optimization: Within a single task, the agent generates an initial output, compares it against a defined standard or feedback from others, and repeatedly refines it in following reasoning steps, until is met. c) Interactive Learning : The agent interacts with a dynamic environment, where experiences can update reusable strategies and, when permitted, high-level goals. Goal adjustment is optional; experience reuse can also support a fixed task goal. This image illustrates three core paradigms for self-improvement in an LLM-based agent: Reflection, Iterative Optimization, and Interactive Learning.
Figure 5. Tool-based reasoning frameworks of LLM-based agent. a) Tool integration studies how to incorporate tools into the agent’s reasoning process; b) Tool Selection addresses which tool from the toolkit is most suitable for the current task or sub-problem; c) Tool Utilization concerns how to effectively operate the chosen tool to generate the desired output. This image is a diagram that illustrates the three main components of a Tool-based reasoning framework for an LLM-based agent. The first component, Tool Integration, shows how tools can be incorporated into an agent's reasoning process through APIs, Plugins, or Middleware. The second, Tool Selection, describes the decision-making process for choosing the most suitable tool from a toolkit, using methods that are either automated, rule-based, or learning-based. The third, Tool Utilization, details how the selected tools are operated to generate a desired output, with examples of sequential, parallel, and iterative execution.
Figure 6. A taxonomy of Multi-agent reasoning frameworks, categorized in a) Organizational Architectures : we summarize three paradigms of multi-agent frameworks to explain how such system can be organized to solve different kinds of challenges; b) Individual Interactions : we demonstrated three types of interaction between different agents within one multi-agent system. This image is a taxonomy of multi-agent reasoning frameworks, divided into two primary categories: Organizational Architectures and Individual Interactions.
| Evidence source | How evidence enters reasoning | Works |
| Literature and reference knowledge | Retrieved passages supply reference knowledge to the reasoning context. | MOOSE-Chem ( Yang et al., 2025d ) ; Stewart et al. ( Stewart and Buehler, 2025 ) ; BioAgents ( Mehandru et al., 2025 ) ; CRISPR-GPT ( Qu et al., 2026 ) ; BioResearcher ( Luo et al., 2025b ) ; DrugAgent ( Inoue et al., 2024 ) . |
| Scientific records and knowledge graphs | Queried records and graph relations are combined as task-specific evidence. | ChatDrug ( Liu et al., 2024b ) , CLADD ( Lee et al., 2026a ) , DrugAgent ( Inoue et al., 2024 ) ; Chemist-X ( Chen et al., 2023 ) ; ChatMOF ( Kang and Kim, 2024 ) , LLaMP ( Chiang et al., 2025 ) . |
| Computational and simulation outputs | Tool outputs provide observations for revising intermediate reasoning and actions. | LIDDIA ( Averly et al., 2025 ) , DrugAgent ( Inoue et al., 2024 ) ; ChatChemTS ( Ishida et al., 2025 ) ; ChemCrow ( M. Bran et al., 2024 ) , CACTUS ( McNaughton et al., 2024 ) ; El Agente Q ( Zou et al., 2025 ) ; AtomAgents ( Ghafarollahi and Buehler, 2025a ) ; ProtAgents ( Ghafarollahi and Buehler, 2024 ) , TourSynbio-Search ( Liu et al., 2024d ) . |
| Execution and workflow feedback | Intermediate results and execution checks guide revision and stage handoffs. | AutoBA ( Zhou et al., 2024a ) ; BioMaster ( Su et al., 2026b ) ; ChemHTS ( Li et al., 2025f ) ; PharmAgents ( Gao et al., 2025b ) , Ruan et al. ( Ruan et al., 2024b ) . |
| Previous experimental observations | Retained observations inform reasoning in subsequent research rounds. | BioDiscoveryAgent ( Roohani et al., 2025 ) . |
Table 4. Evidence access in biochemistry and material science. The middle column summarizes illustrative routes into reasoning; the final column indexes related work. Sources can overlap within a system. Table 5 separately lists evaluation resources.
Figure 8. The BioDiscoveryAgent pipeline ( Roohani et al., 2025 ) , interpreted using our taxonomy. Experimental outcomes inform context updates and subsequent experiment selection. Adapted from ( Roohani et al., 2025 ) . A flowchart of BioDiscoveryAgent showing iterative experimental design and context updates from observed outcomes.
| Metrics Level | ||
| Focus | Related Work | Metrics |
| Lab-involved Biomedical Research | BioResearcher ( Luo et al., 2025b ) | Completeness, Level of Detail, Correctness, Logical Soundness, Structural Soundness |
| High-fidelity Materials Knowledge Retrieval | LLaMP ( Chiang et al., 2025 ) | Precision, coefficient of precision, confidence, self-consistency of response, MAE |
| Drug Discovery | LIDDIA ( Averly et al., 2025 ) | drug-likeness ( Bickerton et al., 2012 ) , Lipinski’s Rule of Five ( Lipinski et al., 1997 ) , synthetic accessibility ( Ertl and Schuffenhauer, 2009 ) , binding affinities ( Trott and Olson, 2010 ) |
| Benchmark/Dataset Level | ||
| Focus | Related Work | Benchmark/Dataset |
Table 5. Evaluation strategies in biochemistry and material science. Metrics assess response or property quality; datasets support repeatable task evaluation; case studies examine complete research workflows. The categories provide complementary evidence.
| Metrics Level | ||||
| Focus | Related Work | Metrics | ||
| AutoSurvey ( Wang et al., 2024h ) | Survey Creation Speed, Content Quality | |||
| Gao et al. ( Gao et al., 2023b ) | Citation Quality | |||
| SurveyX ( Liang et al., 2025 ) | Insertion over Union, semantic-based reference relevance | |||
| SurveyForge ( Yan et al., 2025 ) | Reference, Outline, and Content Quality (SAM Metrics) | |||
| Agent Laboratory ( Schmidgall et al., 2025 ) | Inference cost, Inference time, Success Rate | |||
Table 6. Evaluation strategies for general research agents: output metrics, benchmark and dataset studies, research case studies, and model or human judgments. Corpus sizes describe the listed resources, rather than a common number of evaluated tasks.
| Benchmark/Dataset Level | |||
| Focus | Related Work | Benchmark/Dataset | |
| Clinical Consultation Flow | Wang et al. ( Wang et al., 2025e ) | MVME ( Fan et al., 2025 ) | |
| Zero-shot Medical Reasoning | MedAgents ( Tang et al., 2024 ) | Jin et al. ( Jin et al., 2021 ) ,Pal et al. ( Pal et al., 2022 ) ,PubMedQA ( Jin et al., 2019 ) , Hendrycks et al. ( Hendrycks et al., 2021a ) | |
| Automated supervision of Healthcare Safety | TAO ( Kim et al., 2025b ) | Safetybench ( Zhang et al., 2024i ) ,Medsafetybench ( Han et al., 2024b ) , Chang et al. ( Chang et al., 2025 ) , Hu et al. ( Hu et al., 2024a ) ,Wang et al. ( Wang et al., 2025f ) | |
| Evolvable Medical Agents | MDAgents ( Kim et al., 2024c ) | MedQA ( Jin et al., 2021 ) , PubMedQA ( Jin et al., 2019 ) ,MedBullets ( Chen et al., 2025a ) , JAMA ( Chen et al., 2025a ) ,DDXPlus ( Tchango et al., 2022 ) ,SymCat ( Al-Ars et al., 2023 ) , PathVQA ( He et al., 2020 ) ,PMC-VQA ( Zhang et al., 2024d ) ,MedVidQA ( Gupta et al., 2023 ) ,MIMIC-CXR ( Bae et al., 2023 ) | |
| Healthcare Intent Awareness | MedAide ( Yang et al., 2026 ) | Pre-Diagnosis, Diagnosis, Medicament, Post-Diagnosis Bench ( Yang et al., 2026 ) | |
Table 7. Evaluation strategies for healthcare agents, grouped into dataset tests, system-level evaluation, simulated environments and case studies. Clinical validity requires evidence beyond benchmark or simulation performance.
Figure 9. AMIE’s self-play and inference workflow ( Tu et al., 2025 ) . The inner loop refines simulated dialogues using critic feedback; the outer loop incorporates those dialogues into subsequent fine-tuning. At inference time, the system uses the conversation context and a chain-of-reasoning strategy to construct its next response. Adapted from Tu et al. ( Tu et al., 2025 ) . AMIE uses an inner loop of simulated patient-doctor dialogue with critic feedback and an outer loop of model fine-tuning. The right-hand panel shows the separate inference-time conversation workflow.
| Method / model | Source | HE ( Chen et al., 2021 ) | HE-ET ( Dong et al., 2025 ) | HE+ ( Liu et al., 2023a ) | MBPP ( Austin et al., 2021 ) | MBPP-ET ( Dong et al., 2025 ) | MBPP+ ( Liu et al., 2023a ) | DS-1000 ( Lai et al., 2023 ) | LCB S112 ( Jain et al., 2025 ) | LCB V1 ( Jain et al., 2025 ) |
| Direct baselines | ||||||||||
| GPT-3.5 | ( Huang et al., 2023b ) | 57.3 | 42.7 | – | 52.2 | 36.8 | – | – | – | – |
| GPT-3.5 | ( Islam et al., 2024 ) | 48.1 | 37.2 | 66.5 | 49.8 | 37.7 | – | – | – | – |
| GPT-4 | ( Huang et al., 2023b ) | 67.6 | 50.6 | – | 68.3 | 52.2 | – | – | – | – |
| GPT-4 | ( Islam et al., 2024 ) | 80.1 | 73.8 | 81.7 | 81.1 | 54.7 | – | – | – | – |
| GPT-4o | ( Anthropic, 2024 ) | 90.2 | – | – | – | – | – | – | – | – |
Table 8. Code-generation performance (Pass@1, %). Direct baselines appear first, followed by methods grouped by task model. Source identifies the paper or model report providing the data and experimental design; “–” denotes an unreported result.
| Defects4J v1.2 ( Just et al., 2014 ) | |||
| Method | Source | Loc. | Repair |
| GPT-3.5 | |||
| AgentFL | ( Qin et al., 2025a ) | 157/395 | – |
| ChatGPT + Ochiai | ( Qin et al., 2025a ) | 121/395 | – |
| RepairAgent | ( Bouzenia et al., 2025 ) | – | 74/395 |
| SWE-bench ( Jimenez et al., 2024 ) | |||
Table 9. Software-maintenance results, grouped by task model. Source identifies the reporting study. Defects4J reports successful/evaluated bugs; SWE-bench reports resolved issues (%). Settings differ across reports.
Figure 10. The social-simulation paradigm based on an agentic reasoning framework. (a) Individual (purple, center): Each agent has an initial profile, personal goals, tools and a context that changes as it observes and acts. (b) Interaction with other individuals (blue, left): The agent observes other individuals and interaction events, then reflects on them to update its context and subsequent decisions. (c) Interaction with the environment (green, right): Actions enter a public action space, where they can trigger public events or leave messages; status updates return the resulting changes to the agent. Together, the social and environmental feedback loops allow agents with distinct goals and contexts to behave differently under the same conditions and generate collective dynamics for evaluation. This image presents the fundamental paradigm for social simulation using an Agentic Reasoning Framework. The diagram illustrates a central Individual agent with a profile, personal goals, and tools, which interacts with both Other Individuals and Environments.
| Related work | Simulation focus | Data source | # |
| GOVSIM ( Piatti et al., 2024 ) | Sustainable cooperation | Simulated | – |
| MetaAgents ( Li et al., 2025c ) | Job fair | Simulated | – |
| Generative Agents ( Park et al., 2023 ) | Interactive human behavior | Simulated | 25 |
| RecAgent ( Wang et al., 2025b ) | Recommendation system | Simulated | 20 |
| BotSim ( Qiao et al., 2025 ) | Malicious social botnet | Simulated | 3k |
| S3 ( Gao et al., 2023a ) | Population-level interaction | Simulated | 10k |
Table 10. Social simulation methods and their reported scale. # denotes the reported number of agents or users, retaining each study’s original unit (simulated agents, accounts or source-user pools). k = , M = ; an en dash indicates an unspecified count.
| Method | Model | Source | TB 2.0 ( Merrill et al., 2026 ) | TB 4.0 ( Marten, 2026 ) | GAIA ( Mialon et al., 2024 ) | BrowseComp ( Wei et al., 2025b ) | HLE ( Center for AI Safety et al., 2026 ) | xBench-DS ( Chen et al., 2025b ) | ALE ( Sun et al., 2026 ) |
| FRIDAY / OS-Copilot ( Wu et al., 2024b ) | GPT-4-turbo-1106 | ( Wu et al., 2024b ) | – | – | 40.86/20.13/6.12/- | – | – | – | – |
| AutoGen ( Wu et al., 2024a ) | GPT-4o-2024-08-06 | ( Lu et al., 2026b ) | – | – | -/-/-/6.3 § | – | – | – | – |
| OctoTools ( Lu et al., 2026b ) | GPT-4o-2024-08-06 | ( Lu et al., 2026b ) | – | – | -/-/-/18.4 § | – | – | – | – |
| AutoAgent ( Tang et al., 2026b ) | 3.5 Sonnet | ( Tang et al., 2026b ) | – | – | 71.70/53.49/26.92/55.15 | – | – | – | – |
| DeepVerifier ( Wan et al., 2026 ) | 3.7 Sonnet | ( Wan et al., 2026 ) | – | – | -/-/-/58.93 | 9.0 | – | 44.0 | – |
| Workforce / OWL ( Hu et al., 2025a ) | 3.7 Sonnet | ( Hu et al., 2025a ) | – | – | 84.91/68.60/42.31/69.70 | – | – | – | – |
Table 11. General-purpose agent systems, grouped by model in release order. Composite configurations follow the primary model; undated or undisclosed models appear last. Scores are percentages; GAIA and HLE prioritize full-set over text-only results. GAIA reports level 1/level 2/level 3/avg; ALE reports Near-Term/Full-Spectrum/Last-Exam/overall full-pass rates. Overall scores are source-reported; “-” denotes an unreported component. Source identifies the paper or official resource reporting the data and experimental design.
| Method | Task model | Source | Split | TB * ( Merrill et al., 2026 ) | SWE-V ( OpenAI, 2024 ) | SWE-Pro ( Deng et al., 2025 ) | Gaia2-mini ( Froger et al., 2026 ) | AppWorld ( Trivedi et al., 2024 ) | ARC-AGI-2 ( ARC Prize Foundation, 2025 ) |
| GEPA ( Agrawal et al., 2026 ) | GPT-5 mini | ( Chen et al., 2026b ) | T | 17.6–19.6 | 45.3–46.7 | – | – | 36.7–37.4 | – |
| HarnessFix ( Chen et al., 2026b ) | GPT-5 mini | ( Chen et al., 2026b ) | T | 17.6–26.5 | 45.3–57.3 | – | – | 36.7–43.0 | – |
| Meta-Harness ( Lee et al., 2026b ) | GPT-5 mini | ( Chen et al., 2026b ) | T | 17.6–23.5 | 45.3–54.7 | – | – | 36.7–40.4 | – |
| ReCreate ( Hao et al., 2026 ) | GPT-5 mini | ( Chen et al., 2026b ) | T | 17.6–21.6 | 45.3–51.7 | – | – | 36.7–39.3 | – |
| SCOPE ( Pei et al., 2025 ) | GPT-5 mini | ( Chen et al., 2026b ) | T | 17.6–20.6 | 45.3–48.3 | – | – | 36.7–38.9 | – |
| HarnessFix ( Chen et al., 2026b ) | Sonnet 4.5 | ( Chen et al., 2026b ) | T | 34.3–47.1 | 61.3–71.7 | – | – | 61.1–70.4 | – |
Table 12. Self-evolving systems and harness optimizers (scores in %). Source identifies the reporting study. – : baseline/control and method scores; single values: final scores; “–”: unreported; : independent reference comparison. *: TB = Terminal-Bench 2.0; : TB version unspecified by the source. SWE-V/Pro: SWE-bench Verified/Pro. Evaluation split: O: optimization/search tasks; G: held-out promotion-gate tasks, hidden from the proposer; T: test tasks excluded from search and selection; M: task-weighted total over search and held-out tasks; X: frozen cross-model transfer; U: split unspecified; “cell”: per-score O/T labels.
| Method | Task model | Source | Split | GAIA ( Mialon et al., 2024 ) | HLE ( Center for AI Safety et al., 2026 ) | ALFWorld ( Shridhar et al., 2021 ) | WebShop ( Yao et al., 2022 ) |
| AFlow ( Zhang et al., 2025c ) | GPT-5-Chat | ( Cheng et al., 2026 ) | U | 18.47–19.75 | – | 86.87–93.40 | 25.10–37.90 |
| Alita ( Qiu et al., 2025b ) | GPT-5-Chat | ( Cheng et al., 2026 ) | U | 18.47–72.73 | – | 86.87–86.13 | 25.10–30.21 |
| Mem 2 Evolve ( Cheng et al., 2026 ) | GPT-5-Chat | ( Cheng et al., 2026 ) | U | 18.47–76.31 | – | 86.87–94.31 | 25.10–39.20 |
| GEPA ( Agrawal et al., 2026 ) | GPT-5 mini | ( Chen et al., 2026b ) | T | 43.3–46.7 | – | – | – |
| HarnessFix ( Chen et al., 2026b ) | GPT-5 mini | ( Chen et al., 2026b ) | T | 43.3–61.7 | – | – | – |
| Meta-Harness ( Lee et al., 2026b ) | GPT-5 mini | ( Chen et al., 2026b ) | T | 43.3–56.7 | – | – | – |
Table 12. Self-evolving systems and harness optimizers (continued).
| Scenario | Task constraints | Organization | Tool use | Oversight |
| Scientific research | Computational checks; experimental feedback | Tool loops; specialist review | Retrieval and computation | Goal setting; proposal review |
| Healthcare | Incomplete evidence; high cost of error | Specialist teams; deliberation | Records and diagnostic tools | Tiered and clinician review |
| Software engineering | Artifact dependencies; executable feedback | Controllers or teams; artifact handoffs | Execution and repair search | Sandboxes; independent tests |
| Social/economic simulation | Private objectives; interacting behavior | Individuals or groups; cooperation/competition | Memory and environment tools | Behavioral validation; risk control |
| General-purpose agents | Varied tasks; uneven feedback | Planner–worker teams; verification | Search, code and reusable skills | Permissions; update gates |
Table 13. Task constraints and framework choices across scenarios. Entries summarize conditional patterns; supporting examples and within-scenario differences are discussed in Sections 4.1 – 4.5 .